Event Horizon Updates

DAO Programs & Initiatives
,
6

UPDATE: Transition to Agentic Governance

Users want to be heard, yet they don’t want to ‘do governance’. This is one of the lessons we’ve learned through both our successes and failures thus far (mid-program report coming soon).

When Arbitrum DAO voted to support Event Horizon’s “Voter Enfranchisement Pool” our mandate was clear: increase the number of people who can have their voices heard in Arbitrum DAO. With this learning in mind, Event Horizon is working to build a more intuitive, user-friendly governance UX through what we’ve defined as the Emperor-Consul AI Agent model.

Each voter is assigned a trusted, agentic consul to which they are the emperor. The human voter simply communicates their higher-level desires to their agentic consul, and the consul goes out into the digital agora and continues to represent those interests until told otherwise. This modality compresses the highly taxing governance process detailed below, into simple chat engagements with an AI representative.

Along the way, the consul agents will provide the human voters with reports of their voting behavior, the latest relevant news and events it has come across within each DAO of interest, and more. Through these communications, the consul can be further refined to reinforce actions that please the human voter, and deprecate those which deviate from how the human voter would have otherwise manually voted. The user can also override any decision by voting manually at any time.

In the first couple weeks since the launch of AI agents, we have seen over 120 individual agents created, each with a unique persona, and the user dialog has begun rolling in, further evolving each agent.

Proxy vs Independent Actor: We believe that it is critical to ensure that AI Agents are underpinned by human actors which they represent. The desired outcome is not to have thousands of independently acting AI agents deciding the fate of today’s protocols. Rather it is to use AI agents as proxies for true human desires. In this regard, the creation and inclusion of AI agents must be on a per-person or per-organizational basis.

1. Fixing a Broken Industry UX – The current governance UX standard is broken.
Prior to the creation of Event Horizon, the average voter had no reason to participate in governance whether manually or agentically as they had no voting power to leverage. Every DAO requires millions of dollars to have a meaningful impact. Such voting power is a luxury previously relegated to fewer than 50 individuals across all major DAOs, and of those 50 it was an opportunity typically utilized by about 20.

Before making governance better, Event Horizon had to make it accessible through the creation of the voter pool.

However, in the process of making voting meaningful, Event Horizon made many of the pain points of governance more apparent. By bringing dozens of proposals across 8 DAOs under the same UI umbrella, Event Horizon unintentionally highlighted just how much friction remained in the standard DAO governance UX by requesting the user go through the process tens of times per week.

The average proposal requires weeks to months of rigorous, often fruitless effort. It’s important to place oneself in the mind of a retail voter with little net voting power, no immediate incentive to participate, and limited connection to delegates or the broader governing body. At present, for maximal participation, we are asking this average retail voter to engage as follows:

  • Forum: a voter must read dozens of comments across half a dozen or more proposals over the course of weeks of dialog in the governance forum.
  • Working Groups: sometimes, this process is further fragmented across working groups which then turns the stream of communication found in the forum into a torrent of constant chatter across dozens of time zones in hard-to-follow telegram groups.
  • Video Calls: along with chat communications there is an expectation that a fully informed voter would join multiple hours of Zoom or Meets sessions.
  • Temperature Check: After the communications have been completed the proposal may move to temperature check at which point the user should manually sign their support. And, while it may sound like a small inconvenience when simply written here, for some hardware wallets add further friction. The added effort of finding, connecting, and interacting with a physical device for an already fairly low-reward endeavor does add more barriers to participation. For this effort, the proposal may fail at this point setting the process back to step one.
  • Onchain Proposal: If all the above has been completed, the manual signature process of the temperature check is repeated on chain with the same inconveniences and potential for failure due to rejection or quorum.

In fact, it is a process so ridden with time, friction, setbacks and obstacles that nearly every DAO has had to implement some form of Delegate Incentive Program to encourage participation even amongst the most influential of voters. And, while being a delegate is a valuable contribution to the ecosystem and the difficulty found in being a delegate is aptly compensated such that participation and productivity do occur, this simply doesn’t translate well to the average voter.

Solution: Through agentic governance, Event Horizon is able to eliminate the vast majority of these hurdles. Text-based communication (forum, working group, and even video transcripts) can be fed directly into the model. The temperature check and on-chain votes can be automatically handled for the user by their agent.

2. Greater Voter Inclusion: While it is likely impossible to create an agentic model which represents its underpinning human’s interest with 100% accuracy, narrowing the delta between the would-be manually selected choice of the human actor and the automatically selected choice of the AI agent is crucial. The closer the difference converges, the stronger the case for synthetic decision trees becomes. I.e. 1,000 synthetic agents representing 1,000 retail users with 95% accuracy is a clearly superior structure to 10 human voters representing themselves with 100% accuracy. There is no inherent virtue in human action itself, so long as AI models accurately represent what the human actor would have done, and reducing the friction to participation on the human serves only to increase net representation within the DAO, which has been our mandate from the start.

3. Future Evolution Capacity: AI governance allows for significant future expansion which we will discuss at greater length in a future post.

4. Anti-Sybil: We’ve been reviewing the Sybil behavior noted last week. It’s disappointing that someone, likely within the delegate, channel would use the pool for nothing more than to satisfy their own desire to troll. That said, we have already learned from this instance and implemented the first step of a multi-part solution:

  • Align Around Agent Voting: We have consolidated all voting around agent interactions. One can still manually tell their agent to vote in specific ways, but the actual voting is conducted by the agent (which we can directly injuncted upon if need be). Absent agents providing a point of intervention, there is a difficult balance between sybil prevention and user friction which occurs when you require added proof of humanity. Scaling up individual wallet sybil requirements blocks trolls, but also adds friction to real users and stifles our mandate to drive higher participation.

  • IP tracking: Because we are taking a more interventionist approach, we can use IP access as a sybil filter. Event Horizon will purchase and blacklist all major VPN addresses and track IP-based engagement to internally identify clusters. If behavior is suspicious, Event Horizon will freeze agent clusters and provide a report and rationale to the DAO. Any affected wallets would also be able to petition their blacklist here should they wish to come public with their ownership of any potentially erroneously blocked agents.

  • Conversational Monitoring: All voting engagements will require the provision of a request or preference set to the AI agent in the form of conversation. Event Horizon will have access to all chat logs with the agents which will further allow for the identification of similarities between the reasonings / intentions passed through to the agents.

  • Increasing Net Participation: As the agent count increases, the threshold for Sybil capture rises. Since transitioning to fully agentic voting, the voter turnout has already risen from ~30 during the Sybil instance this past Friday to over 100 and rising as each new agent comes online. In the coming weeks, we expect >250 new agents. In the coming months, we expect to cross 1,000 with each agent further hardening the system against individual Sybil behavior.

  • (non-AI Sybil Improvement) Improvements to MSS timing and Notification: Event Horizon will expand the MSS veto window to extend 22 hours after the EH snapshot closure and will notify MSS members of proposals that closed with suspicious contexts including large end-of-vote influx and close margin outcomes.

We believe that this direct involvement will better protect the pool and the trade-off of greater centralization is justifiable given the pool is a public good. It is a free privilege and not a right. No users paid for this VP and voters are not entitled to it should they engage in suspicious behavior.

5. Greater Rationale Capture: each of the now >100 agents is providing a bespoke rationale set for each proposal in accordance with the preferences of its underpinning human voter. We can already see quite granularly why an agent is voting on behalf of its human.

6. Faster Decision Time: Given all voting is done through agents, the start-to-finish voting process can be compressed down to minutes.

7. Fidelity to Voter Preference: As stated above, Event Horizon is working to compress the delta between the agentic outcome and the otherwise desired outcome of the human voter as greatly as possible. Satisfactory margin can be identified further through post-vote follow-ups with the human voter to check for agent-human alignment and can be further inferred by the rate of continued usage. Should a user choose to continue allowing their agent to operate and vote in their proxy, there is an assumption that the user is satisfied with the degree to which the agent has accurately represented their interest. And, again, at any point, the user may chat with their agent to modify its behavior pattern should they feel the agent isn’t ideally aligned with their interests or should they simply have a change in their interests over time.

8. Alignment with Initial Mandate:

  • Manual Voting is Still Possible: Users will still be able to vote manually as was the core mandate of the original proposal. The only subtle difference is that manual voting will be done by requesting one’s agent vote a specific way on a specific proposal, vs manually selecting yes or no as a button on the proposal UI.

  • Increase participation: Part of our core mandate was to drive higher voting participation. The two greatest barriers to participation are motivation and difficulty. We initially tackled increasing motivation by increasing voting power. However, we learned that this is only half of the equation. If the difficulty is sufficiently high as to exceed motivation, only the absolute apex governance dedicated individuals show up. Imagine effort as a point system. To willingly participate, a users ‘motivation points’ must exceed the ‘difficulty points’ which they perceive. If most users maintain 30 points of motivation in governance, but face 100 points of difficulty, they will not show up. We erroneously focused solely on increasing motivational points. When in fact, the difficulty barrier was quite high and increasing motivation was only half of the issue. But if instead, we can also lower the difficulty to 10 points via agentic simplification and automation vs manual engagement and repeated user obligations, anyone with greater than 10 points of motivation should enter the ecosystem. It’s a two-sided issue and we believe strongly this will address the to-date, unaddressed half of the equation.

Ultimately, AI and Agentic structures are a when, not if, for governance. And, Event Horizon sees it as imperative that it be done with the interests of the existing ecosystem stakeholders (delegates, users, foundation, and more) considered vs at the inevitable rogue swarm alternative. Event Horizon is here to bring Agentic governance to light while working hand-in-hand with each of you here. Please feel free to contribute suggestion and feedback below hopefully in the spirit of teamwork, collaboration, and the best interest of the ecosystem as our north star.